Tencent's Gander separates real-time chat from background reasoning

Tencent's Hunyuan Speech team, working with researchers at several universities, has introduced Gander, a research model built to hold a real-time conversation while a separate component works through complex tasks in the background. Gander processes speech, video, and text at the same time, and a user can interrupt it at any point; the model can also ask follow-up questions or offer progress updates on its own.

The design splits work between two roles the researchers call the cerebellum and the brain, borrowing the terms from human anatomy. The cerebellum handles the moment-to-moment conversation: it breaks the exchange into one-second segments and decides, without a separate voice-activity detector, when to listen, speak, or stop because the user has interrupted, using roughly the last two minutes of the conversation as memory. The brain handles reasoning and complex tasks, and can be swapped for agent systems such as Codex or Claude Code without retraining the conversation model; in Tencent's tests, an unspecified model from OpenAI's GPT-5.6 family filled that role.

Tested on Full-Duplex-Bench v3, a benchmark that scores voice assistants across different task scenarios, Gander started speaking at the right moment in all 100 scenarios and interrupted users in 8 percent of cases, against 13.5 percent for GPT-Realtime and nearly 48 percent for the weakest competitor tested. The researchers say Gander uses a relatively small model to compete with commercial systems including GPT-Realtime, Gemini Live, and Grok.

Task accuracy is weaker: the benchmark scores the whole system, so speech recognition and output errors count against Gander, and the brain performs much better when given text directly. Video and audio understanding also suffered, with Gander scoring worse than its base model on one test involving counting and locating objects in images; the researchers attribute this to training that favors fluid conversation over precise perception.

Gander was trained on about 2.7 million examples, some of which teach it to stay quiet when there is background noise or when nobody in a group conversation is addressing it. The team plans to publish the model weights and training data once it completes what it calls "the open source release process"; a GitHub repository for the code already exists, with demos on the project page.

Gander follows Tencent's July release of Hy3, an open language model the company says narrowed the gap with rivals, especially on agent tasks; Hy3 already runs in WorkBuddy, Yuanbao, and WeChat. Tencent is also negotiating to take the largest stake in agent startup Manus after Beijing blocked Meta's attempt to acquire it, a move the company sees as fitting its own plans, including an agent embedded in WeChat.

Gander is part of a broader trend of splitting agent work across models: OpenAI's GPT-Live separates conversation from reasoning, handing web searches and agent tasks to a background model while the chat continues; Sakana AI's Fugu is a separate language model that calls others from an expandable pool; and OpenAI is testing proactive agents that create follow-up tasks and contact users unprompted. An Anthropic analysis cited in the piece found that experienced users interrupt Claude Code in about 9 percent of work steps, versus roughly 5 percent for newcomers, and a survey found that teams building conversational voice and chat agents report latency problems particularly often.

Key facts

Gander splits work between a "cerebellum" that manages real-time conversation and a swappable "brain" that handles complex tasks; in Tencent's tests the brain role was filled by an unspecified model from OpenAI's GPT-5.6 family.
On Full-Duplex-Bench v3, Gander started speaking at the right moment in all 100 scenarios and interrupted users in 8 percent of cases, versus 13.5 percent for GPT-Realtime and nearly 48 percent for the weakest competitor tested.
Gander trails on task accuracy and shows weaker video and audio understanding than its own base model, which the researchers blame on training that favors conversational fluidity over precise perception.
The model was trained on about 2.7 million examples; Tencent plans to release the weights and training data once it completes its open source release process, and a GitHub repository for the code already exists.
Gander follows Tencent's July release of the open model Hy3 and comes as Tencent negotiates to take the largest stake in agent startup Manus after Beijing blocked a Meta acquisition attempt.

Why it matters

Today's voice assistants mostly take turns with users, the researchers argue, while real conversation involves interruption, quick feedback, and listening while talking. Gander is Tencent's attempt to close that gap by continuously processing speech, video, and text and letting a separate background component handle anything that needs real planning time, so the conversational half never has to slow down to think. It fits a wider pattern: OpenAI's GPT-Live and Sakana AI's Fugu both split conversation or orchestration from heavier reasoning across separate models, suggesting the industry sees single-model voice assistants as hitting a speed-versus-reasoning ceiling.

Who it affects

The direct audience is researchers and engineers building conversational voice and agent systems, who now have a published architecture and benchmark numbers to compare against. It also touches Tencent's own product line: the sibling model Hy3 already runs inside WorkBuddy, Yuanbao, and WeChat, and Tencent's push to take the largest stake in agent startup Manus points to where a production version of this cerebellum-and-brain split could end up. Competitors named in the benchmark, including the teams behind GPT-Realtime, Gemini Live, and Grok, are also affected as a direct point of comparison.

How to use it

Gander is a research model, not a shipped product. The team has not released the weights or training data yet; it says publication will follow completion of "the open source release process," with no date given. A GitHub repository for the code already exists, with demos posted on the project page, so the architecture and some materials can be inspected now even though the trained model itself is not yet available.

How solid is it

There is no established benchmark for models that combine conversation and background agent work, so the researchers evaluated Gander on Full-Duplex-Bench v3, a benchmark built for voice assistants generally. Gander's lead is specific and measured: 100 out of 100 correctly timed responses and an 8 percent interruption rate, against 13.5 percent for GPT-Realtime and about 48 percent for the weakest system tested. But the same report shows the tradeoff plainly: task accuracy lags because the benchmark scores the full pipeline, including speech recognition errors, and the brain component does much better when it receives text directly rather than through Gander's conversational front end.

Risks and caveats

The researchers themselves call the work early: how to scale Gander up is an open question, and there is no standard way yet to evaluate systems built this way. Video and audio understanding regressed compared with Gander's own base model, which the team attributes to training that favors smooth conversation over precise perception of what is on screen or in the audio. The brain component used in testing was only described as "an unspecified model from OpenAI's GPT-5.6 family," which limits how precisely outside teams can reproduce the reported numbers. No parameter count, architecture size, or compute budget is disclosed, and no release date has been set for the promised weights and training data.